The aging of transportation infrastructure and large production facilities has become a serious problem, leading to an increased demand for robotic work support. To meet this demand, robots need to be installed in environments that are difficult to access due to a lack of footholds, such as high and narrow places, but this is difficult. To solve this problem, we have developed ``ManipuRailer'', a robot arm comprising linear and rotational joints and equipped with multiple rails for robot locomotion and support. A scale model is fabricated, and the followings are clarified through experiments. To prevent torsional deformation, it is effective to design a symmetrical rail structure with all arm links and joints placed on a single plane. The following three design and operational innovations are effective in reducing the driving torque of long reaching arm. (a1) The parts of the sliding mechanism should be installed at the leading edge of the rotating part. (a2) The pully and timing belt should be placed so that the sprocket and tooth surface are near the restraining surface by the guide rail. (b1) The center of gravity of the rotating body is shifted by the rail sliding mechanism. These measures shorten the distance between the rail and the axis of rotation and reduce the load moment.

PYNet has imporoved the accuracy of pose estimation for simple-shaped objects by initially estimating the object's grounding surface. This approach reduces the degrees of freedom in the pose to be estimated. However, estimating poses that are 180° rotated remains a challenging task. To address this problem, we developed a new method that also estimates the height direction of the object. This additional constraint on the pose of the simple-shaped object being estimated further improves the pose estimation accuracy. When compared to the conventional PYNet, our method achieved an accuracy improvement of 2.41pt in the original RGBD camera dataset and 2.78pt in the pubic YCB object dataset.

This research proposes a method using Large Language Model (LLM) to derive appropriate text prompts for image recognition tasks under various conditions. This method leverages In-context Learning and Few-shot Prompting to enable LLM to understand tasks based on provided data and execute them with minimal examples. Experiments on image recognition in logistics robot picking scenes confirmed the method's effectiveness in deriving suitable prompts for diverse scenarios, multi-product situations and including shrink-wrapped items.

A self-localization system that can switch seamlessly positions and attitudes estimated by a NDT scan matching and a RTK-GNSS has been developed so far. In this paper, a 3D urban model provided by the Project PLATEAU is utilized as a 3D environmental map for our seamless switchable self-localization system. An autonomous driving experiment is conducted using this self-localization system with a 3D urban model in the Tsukuba Challenge 2023EX, and the mobile robot completed a designated course involving around 1[km] in an urban area. However, one of issues encountered by using a 3D urban model of the Project PLATEAU is the jump in estimated positions when switching from a NDT scan matching to a RTK-GNSS. This issue can result in meandering autonomous driving. Thus, methods to solve this problem are discussed and proposed. Consequently, the methods enable mobile robots to stably obtain seamless estimated positions and attitudes.

This paper focuses on object storing tasks by a dual-arm robot. The robot is required to store objects into empty spaces in a shelf. However, the empty spaces vary according to the positions of objects in the shelf. For the similar images with the objects, we propose an empty space classifier based on CNN. The CNN is trained through deep metric learning. From the classifier, empty space images are generated. These images are further used as inputs to a motion planner. In the experiments, it is shown that the robot is able to store objects into the correct empty spaces.

Considering the coexistence of robot and human, robots must avoid colliding with human for safety. Therefore, we developed high-speed skeleton tracking method with deep learning-based detection and optical flow. Furthermore, we also developed real-time collision avoidance path planning method dealing with dynamic human arm skeleton as obstacles based on velocity potential field. We achieved tracking 6 joints of arms with 380[Hz], more than 5 times faster than the previous method. We also verified effectiveness of our collision avoidance planning method in simulation and confirmed that it works with under 100 microseconds.

Cooperative exploration by unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs) is efficient for long-duration surveys and wider exploration areas. We propose a rendezvous point determination and path planning method that reduces the travel distance between survey points in a UAV-UGV cooperative system, in which the UAV can recharge itself by landing on the UGV. In this study, Bidirectional Informed Rapidly-exploring Random Tree* that selectively extends paths for UAV and UGV in a specific direction is developed. Simulation study in three different terrains confirms that this method can reduce the travel distance by 5–7% compared to existing methods.

In this paper, we describe the transport performance of our proposed device, based on transportation experiments conducted with three types of simulated feces, each exhibiting different physical characteristics according to the Bristol Scale, which evaluates fecal hardness. In long-term space missions, the environment is maintained through regenerative and recycling processes, including the recycling of carbon dioxide, sweat, and other materials. The authors are developing a feces transport device to facilitate the recycling of feces. The results of the transportation experiments with simulated feces of varying Bristol Scale types indicate that the device successfully transported feces regardless of their physical characteristics.